Fixes for v2.2.7, mostly parallelization

[ckronlage]

• Fixes that allow parallelizing MELD-graph inference (no hardcoded temp file paths, wrapper around open_hdf5_file that just retries it the file can't be opened in parallel)
• Add --threads argument for calling freesurfer with multiple threads (also useful for parallelizing)
• --skip-feature-plotting speeds up inference for big cohorts when these plots are not needed
• Move the MELD-graph prediction outputs in surface space etc. to the predictions_reports/ folder, e.g.:

output/predictions_reports/sub-27/predictions
├── prediction.nii.gz
├── lh.prediction.nii.gz
├── rh.prediction.nii.gz
├── fsaverage_sym
│   ├── lh.prediction.mgh
│   └── rh.prediction.mgh
├── surf_native
│   ├── lh.prediction.mgh
│   └── rh.prediction.mgh
├── vol_freesurfer
│   ├── lh.prediction.mgz
│   └── rh.prediction.mgz
└── vol_native
    ├── lh.prediction.mgz
    └── rh.prediction.mgz

Which makes them easier to find and also avoids conflicts when running multiple predictions (with different harmo parameters) using the same freesurfer outputs.

• Write the freesurfer expert-options fix for FLAIRpial segmentation to a per-subject file (instead of relying on the output/freesurfer_outputs directory to contain it)
• Update docs to explain the changes and give an example of how inference can be parallelized for large cohorts

[mathrip]

thanks for the changes, they're great. I've added this PR to the new version v2.2.7 and created a new PR #117